Upload fused decoder-joint model bundle

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+parakeet-v2/onnx/encoder-model.onnx.data filter=lfs diff=lfs merge=lfs -text

parakeet-v2/decoder-joint.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d8bce5600493ba8d0215107bd22acb55fa6f34ecb774b9393550797def21240f
+size 243

parakeet-v2/decoder-joint.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f5234795b3e724daaf73251d9e3f0f1fb91f37783fcf13d5765b439eee9fcba
+size 595

parakeet-v2/decoder-joint.mlmodelc/model.mil

@@ -0,0 +1,98 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios15>(tensor<fp32, [1, 1024, 1]> encoder_outputs, tensor<fp32, [2, 1, 640]> input_states_1, tensor<fp32, [2, 1, 640]> input_states_2, tensor<int32, [1]> target_length, tensor<int32, [1, 1]> targets) {
+            tensor<fp32, [1025, 640]> embedding_weight = const()[name = tensor<string, []>("embedding_weight"), val = tensor<fp32, [1025, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp32, [2560]> layer_1_bias = const()[name = tensor<string, []>("layer_1_bias"), val = tensor<fp32, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2624128)))];
+            tensor<fp32, [2560, 640]> layer_1_hidden_weights = const()[name = tensor<string, []>("layer_1_hidden_weights"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2634432)))];
+            tensor<fp32, [2560, 640]> layer_1_input_weights = const()[name = tensor<string, []>("layer_1_input_weights"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9188096)))];
+            tensor<fp32, [2560]> layer_2_bias = const()[name = tensor<string, []>("layer_2_bias"), val = tensor<fp32, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15741760)))];
+            tensor<fp32, [2560, 640]> layer_2_hidden_weights = const()[name = tensor<string, []>("layer_2_hidden_weights"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15752064)))];
+            tensor<fp32, [2560, 640]> layer_2_input_weights = const()[name = tensor<string, []>("layer_2_input_weights"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22305728)))];
+            tensor<fp32, [640]> encoder_projection_bias = const()[name = tensor<string, []>("encoder_projection_bias"), val = tensor<fp32, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28859392)))];
+            tensor<fp32, [640, 1024]> encoder_projection_weight = const()[name = tensor<string, []>("encoder_projection_weight"), val = tensor<fp32, [640, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28862016)))];
+            tensor<fp32, [640]> prediction_projection_bias = const()[name = tensor<string, []>("prediction_projection_bias"), val = tensor<fp32, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31483520)))];
+            tensor<fp32, [640, 640]> prediction_projection_weight = const()[name = tensor<string, []>("prediction_projection_weight"), val = tensor<fp32, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31486144)))];
+            tensor<fp32, [1030]> joint_projection_bias = const()[name = tensor<string, []>("joint_projection_bias"), val = tensor<fp32, [1030]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33124608)))];
+            tensor<fp32, [1030, 640]> joint_projection_weight = const()[name = tensor<string, []>("joint_projection_weight"), val = tensor<fp32, [1030, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33128832)))];
+            tensor<int32, [2]> var_19_begin_0 = const()[name = tensor<string, []>("op_19_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_19_end_0 = const()[name = tensor<string, []>("op_19_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_19_end_mask_0 = const()[name = tensor<string, []>("op_19_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<bool, [2]> var_19_squeeze_mask_0 = const()[name = tensor<string, []>("op_19_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
+            tensor<int32, [1]> var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, squeeze_mask = var_19_squeeze_mask_0, x = targets)[name = tensor<string, []>("op_19")];
+            tensor<int32, []> inputs_1_axis_0 = const()[name = tensor<string, []>("inputs_1_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, 640]> inputs_1 = gather(axis = inputs_1_axis_0, indices = var_19, x = embedding_weight)[name = tensor<string, []>("inputs_1")];
+            tensor<int32, [3]> hidden_state_1_begin_0 = const()[name = tensor<string, []>("hidden_state_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> hidden_state_1_end_0 = const()[name = tensor<string, []>("hidden_state_1_end_0"), val = tensor<int32, [3]>([1, 1, 640])];
+            tensor<bool, [3]> hidden_state_1_end_mask_0 = const()[name = tensor<string, []>("hidden_state_1_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
+            tensor<bool, [3]> hidden_state_1_squeeze_mask_0 = const()[name = tensor<string, []>("hidden_state_1_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
+            tensor<fp32, [1, 640]> hidden_state_1 = slice_by_index(begin = hidden_state_1_begin_0, end = hidden_state_1_end_0, end_mask = hidden_state_1_end_mask_0, squeeze_mask = hidden_state_1_squeeze_mask_0, x = input_states_1)[name = tensor<string, []>("hidden_state_1")];
+            tensor<int32, [3]> cell_state_1_begin_0 = const()[name = tensor<string, []>("cell_state_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> cell_state_1_end_0 = const()[name = tensor<string, []>("cell_state_1_end_0"), val = tensor<int32, [3]>([1, 1, 640])];
+            tensor<bool, [3]> cell_state_1_end_mask_0 = const()[name = tensor<string, []>("cell_state_1_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
+            tensor<bool, [3]> cell_state_1_squeeze_mask_0 = const()[name = tensor<string, []>("cell_state_1_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
+            tensor<fp32, [1, 640]> cell_state_1 = slice_by_index(begin = cell_state_1_begin_0, end = cell_state_1_end_0, end_mask = cell_state_1_end_mask_0, squeeze_mask = cell_state_1_squeeze_mask_0, x = input_states_2)[name = tensor<string, []>("cell_state_1")];
+            tensor<fp32, [2560]> var_41_bias_0 = const()[name = tensor<string, []>("op_41_bias_0"), val = tensor<fp32, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35765696)))];
+            tensor<fp32, [1, 2560]> var_41 = linear(bias = var_41_bias_0, weight = layer_1_input_weights, x = inputs_1)[name = tensor<string, []>("op_41")];
+            tensor<fp32, [1, 2560]> var_43 = linear(bias = var_41_bias_0, weight = layer_1_hidden_weights, x = hidden_state_1)[name = tensor<string, []>("op_43")];
+            tensor<fp32, [1, 2560]> var_44 = add(x = var_41, y = var_43)[name = tensor<string, []>("op_44")];
+            tensor<fp32, [1, 2560]> gates_1 = add(x = var_44, y = layer_1_bias)[name = tensor<string, []>("gates_1")];
+            tensor<int32, [4]> var_46_split_sizes_0 = const()[name = tensor<string, []>("op_46_split_sizes_0"), val = tensor<int32, [4]>([640, 640, 640, 640])];
+            tensor<int32, []> var_46_axis_0 = const()[name = tensor<string, []>("op_46_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 640]> var_46_0, tensor<fp32, [1, 640]> var_46_1, tensor<fp32, [1, 640]> var_46_2, tensor<fp32, [1, 640]> var_46_3 = split(axis = var_46_axis_0, split_sizes = var_46_split_sizes_0, x = gates_1)[name = tensor<string, []>("op_46")];
+            tensor<fp32, [1, 640]> input_gate_3 = sigmoid(x = var_46_0)[name = tensor<string, []>("input_gate_3")];
+            tensor<fp32, [1, 640]> forget_gate_3 = sigmoid(x = var_46_1)[name = tensor<string, []>("forget_gate_3")];
+            tensor<fp32, [1, 640]> candidate_gate_3 = tanh(x = var_46_2)[name = tensor<string, []>("candidate_gate_3")];
+            tensor<fp32, [1, 640]> output_gate_3 = sigmoid(x = var_46_3)[name = tensor<string, []>("output_gate_3")];
+            tensor<fp32, [1, 640]> var_54 = mul(x = forget_gate_3, y = cell_state_1)[name = tensor<string, []>("op_54")];
+            tensor<fp32, [1, 640]> var_55 = mul(x = input_gate_3, y = candidate_gate_3)[name = tensor<string, []>("op_55")];
+            tensor<fp32, [1, 640]> next_cell_1 = add(x = var_54, y = var_55)[name = tensor<string, []>("next_cell_1")];
+            tensor<fp32, [1, 640]> var_57 = tanh(x = next_cell_1)[name = tensor<string, []>("op_57")];
+            tensor<fp32, [1, 640]> inputs = mul(x = output_gate_3, y = var_57)[name = tensor<string, []>("inputs")];
+            tensor<int32, [3]> hidden_state_begin_0 = const()[name = tensor<string, []>("hidden_state_begin_0"), val = tensor<int32, [3]>([1, 0, 0])];
+            tensor<int32, [3]> hidden_state_end_0 = const()[name = tensor<string, []>("hidden_state_end_0"), val = tensor<int32, [3]>([2, 1, 640])];
+            tensor<bool, [3]> hidden_state_end_mask_0 = const()[name = tensor<string, []>("hidden_state_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
+            tensor<bool, [3]> hidden_state_squeeze_mask_0 = const()[name = tensor<string, []>("hidden_state_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
+            tensor<fp32, [1, 640]> hidden_state = slice_by_index(begin = hidden_state_begin_0, end = hidden_state_end_0, end_mask = hidden_state_end_mask_0, squeeze_mask = hidden_state_squeeze_mask_0, x = input_states_1)[name = tensor<string, []>("hidden_state")];
+            tensor<int32, [3]> cell_state_begin_0 = const()[name = tensor<string, []>("cell_state_begin_0"), val = tensor<int32, [3]>([1, 0, 0])];
+            tensor<int32, [3]> cell_state_end_0 = const()[name = tensor<string, []>("cell_state_end_0"), val = tensor<int32, [3]>([2, 1, 640])];
+            tensor<bool, [3]> cell_state_end_mask_0 = const()[name = tensor<string, []>("cell_state_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
+            tensor<bool, [3]> cell_state_squeeze_mask_0 = const()[name = tensor<string, []>("cell_state_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
+            tensor<fp32, [1, 640]> cell_state = slice_by_index(begin = cell_state_begin_0, end = cell_state_end_0, end_mask = cell_state_end_mask_0, squeeze_mask = cell_state_squeeze_mask_0, x = input_states_2)[name = tensor<string, []>("cell_state")];
+            tensor<fp32, [1, 2560]> var_74 = linear(bias = var_41_bias_0, weight = layer_2_input_weights, x = inputs)[name = tensor<string, []>("op_74")];
+            tensor<fp32, [1, 2560]> var_76 = linear(bias = var_41_bias_0, weight = layer_2_hidden_weights, x = hidden_state)[name = tensor<string, []>("op_76")];
+            tensor<fp32, [1, 2560]> var_77 = add(x = var_74, y = var_76)[name = tensor<string, []>("op_77")];
+            tensor<fp32, [1, 2560]> gates = add(x = var_77, y = layer_2_bias)[name = tensor<string, []>("gates")];
+            tensor<int32, [4]> var_79_split_sizes_0 = const()[name = tensor<string, []>("op_79_split_sizes_0"), val = tensor<int32, [4]>([640, 640, 640, 640])];
+            tensor<int32, []> var_79_axis_0 = const()[name = tensor<string, []>("op_79_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 640]> var_79_0, tensor<fp32, [1, 640]> var_79_1, tensor<fp32, [1, 640]> var_79_2, tensor<fp32, [1, 640]> var_79_3 = split(axis = var_79_axis_0, split_sizes = var_79_split_sizes_0, x = gates)[name = tensor<string, []>("op_79")];
+            tensor<fp32, [1, 640]> input_gate = sigmoid(x = var_79_0)[name = tensor<string, []>("input_gate")];
+            tensor<fp32, [1, 640]> forget_gate = sigmoid(x = var_79_1)[name = tensor<string, []>("forget_gate")];
+            tensor<fp32, [1, 640]> candidate_gate = tanh(x = var_79_2)[name = tensor<string, []>("candidate_gate")];
+            tensor<fp32, [1, 640]> output_gate = sigmoid(x = var_79_3)[name = tensor<string, []>("output_gate")];
+            tensor<fp32, [1, 640]> var_87 = mul(x = forget_gate, y = cell_state)[name = tensor<string, []>("op_87")];
+            tensor<fp32, [1, 640]> var_88 = mul(x = input_gate, y = candidate_gate)[name = tensor<string, []>("op_88")];
+            tensor<fp32, [1, 640]> next_cell = add(x = var_87, y = var_88)[name = tensor<string, []>("next_cell")];
+            tensor<fp32, [1, 640]> var_90 = tanh(x = next_cell)[name = tensor<string, []>("op_90")];
+            tensor<fp32, [1, 640]> layer_2_hidden = mul(x = output_gate, y = var_90)[name = tensor<string, []>("layer_2_hidden")];
+            tensor<int32, []> var_97_axis_0 = const()[name = tensor<string, []>("op_97_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 1, 640]> output_states_1 = stack(axis = var_97_axis_0, values = (inputs, layer_2_hidden))[name = tensor<string, []>("op_97")];
+            tensor<int32, []> var_100_axis_0 = const()[name = tensor<string, []>("op_100_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 1, 640]> output_states_2 = stack(axis = var_100_axis_0, values = (next_cell_1, next_cell))[name = tensor<string, []>("op_100")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_5_axes_0 = const()[name = tensor<string, []>("input_5_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1, 640]> input_5 = expand_dims(axes = input_5_axes_0, x = layer_2_hidden)[name = tensor<string, []>("input_5")];
+            tensor<fp32, [1, 1, 1024]> input_3 = transpose(perm = input_3_perm_0, x = encoder_outputs)[name = tensor<string, []>("transpose_4")];
+            tensor<fp32, [1, 1, 640]> var_108 = linear(bias = encoder_projection_bias, weight = encoder_projection_weight, x = input_3)[name = tensor<string, []>("linear_0")];
+            tensor<int32, [1]> var_110_axes_0 = const()[name = tensor<string, []>("op_110_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1, 1, 640]> var_110 = expand_dims(axes = var_110_axes_0, x = var_108)[name = tensor<string, []>("op_110")];
+            tensor<fp32, [1, 1, 640]> var_113 = linear(bias = prediction_projection_bias, weight = prediction_projection_weight, x = input_5)[name = tensor<string, []>("linear_1")];
+            tensor<int32, [1]> var_115_axes_0 = const()[name = tensor<string, []>("op_115_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1, 1, 640]> var_115 = expand_dims(axes = var_115_axes_0, x = var_113)[name = tensor<string, []>("op_115")];
+            tensor<fp32, [1, 1, 1, 640]> var_117 = add(x = var_110, y = var_115)[name = tensor<string, []>("op_117")];
+            tensor<fp32, [1, 1, 1, 640]> input = relu(x = var_117)[name = tensor<string, []>("input")];
+            tensor<fp32, [1, 1, 1, 1030]> outputs = linear(bias = joint_projection_bias, weight = joint_projection_weight, x = input)[name = tensor<string, []>("linear_2")];
+            tensor<int32, [1]> var_127 = const()[name = tensor<string, []>("op_127"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> prednet_lengths = add(x = target_length, y = var_127)[name = tensor<string, []>("op_129")];
+        } -> (outputs, prednet_lengths, output_states_1, output_states_2);
+}

parakeet-v2/decoder-joint.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b622dfd4e8ab8e680e3297ac1cc76a50a1fd22eb3526fca6d9e5f7a4e8626b0
+size 35776000

parakeet-v2/onnx/decoder_joint-model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cbb52a07bd70ab5b67f8439d4b3cd8704b18467b4430bcacb5adabe154b8d191
+size 35792059

parakeet-v2/onnx/encoder-model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3987bcd28175d829d12888a996a84e8f62a0e374d9ffd640662c1515adc679d3
+size 41770866

parakeet-v2/onnx/encoder-model.onnx.data

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4dab7362d4874d85965045b1e41b2d61dd2cc0fb25671a7f6b3dc47bf120cc41
+size 2435420160

vad/silero-vad.mlmodelc/analytics/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f856f53009d001936ae46da67ba5edb8a3d28d4ff50ed38646d9c025cc5fbb35
+oid sha256:ad29a493052b54b02adb7c1d81721e9440d21304a481d72c0342b8e0f1031172
 size 243

vad/silero-vad.mlmodelc/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e30e28650a03afaa1bd20041362933235b455245468f533c031b683c29730f19
+oid sha256:6d0300f47cef189a491bcba386a7a990303e55805a118f11f47decc9e86a3e23
 size 459

vad/silero-vad.mlmodelc/metadata.json

@@ -70,7 +70,7 @@
       "name" : "MLModelType_mlProgram"
     },
     "userDefinedMetadata" : {
-      "com.github.apple.coremltools.conversion_date" : "2026-02-28",
+      "com.github.apple.coremltools.conversion_date" : "2026-04-03",
       "com.github.apple.coremltools.source" : "torch==2.10.0",
       "com.github.apple.coremltools.version" : "9.0",
       "com.github.apple.coremltools.source_dialect" : "TorchScript"